Introduction: Pump head and flow rate describe two different jobs in a multi-GPU loop, and reading them together explains why real coolant flow depends on the whole circuit.
A specification sheet can list a pump's maximum head and maximum flow side by side, which makes them look like two versions of the same performance number. They are not. One describes the pressure the pump can build; the other describes the volume it can move in a given time. In a loop with several GPU cold plates, long tubing runs, and a radiator, those two numbers meet a third factor: resistance. Following the fluid from the pump outlet, through the cold plates, and back to the reservoir shows how head and flow divide the work. Engineers comparing liquid cooling solutions often start with these two figures, then discover that the installed circuit decides what they mean in practice.
What pump head adds to a multi-GPU loop
Pump head is the pressure-raising capability of the pump, expressed as the height of a liquid column it could support. A pump with 5 m nominal head can raise water about 5 m, or push against an equivalent pressure drop of roughly 49 kPa (0.49 bar). In a closed multi-GPU loop, the fluid returns to the pump, so elevation largely cancels out. What remains is the pressure needed to overcome friction and component losses. Head is therefore the budget the pump provides for pushing coolant through every cold plate, hose, and radiator in the path. The important detail is that a pump does not deliver its maximum head and maximum flow at the same time. At zero flow, it produces its highest pressure, known as shutoff head. As flow increases, the head available at the outlet falls, because the impeller and internal passages become less effective at higher volume rates. The published pump curve captures this trade-off. A loop with one cold plate may sit near the high-flow end of the curve. Add more cold plates in series, and the same pump moves to a lower-flow, higher-head point. The 5 m figure is the ceiling of the pressure budget, not the pressure seen at every operating condition. This is why a multi-GPU loop asks more of head than a single-block loop. Each additional cold plate adds its own pressure drop, and those drops accumulate when coolant passes through them one after another. Even a well-designed cold plate has a pressure cost. Four plates in series can require several times the pressure of one plate at the same flow. The pump must have enough head left after the radiator, tubing, and fittings have taken their share.
How flow rate interacts with cold plate and tubing resistance
Flow rate is the volume of coolant moved per unit time, usually stated in litres per hour. It tells you how much fluid passes a point, not how hard the pump is pushing. The two quantities are linked through the loop's resistance: for a given circuit, higher flow always requires more pressure drop. In the turbulent range typical of cooling loops, pressure drop rises faster than flow, roughly with the square of flow for many fittings and sudden changes in direction. Doubling flow through the same circuit can therefore demand about four times the pressure, which is why more flow is never free. Three features of the loop shape its resistance most.
- Series cold plates. When coolant passes through GPU cold plates one after another, every plate sees the same flow, and their pressure drops add up. Four identical plates in series produce about four times the pressure drop of one at the same flow, so the pump needs more head to maintain the same volume rate. Series routing is simple and gives each plate the full loop flow, but it builds resistance quickly as GPU count rises.
- Parallel branches. In a parallel layout, the total flow splits among branches, and the pressure drop across each branch is the same. Total resistance is lower than a pure series path, but flow does not divide evenly unless the branches are hydraulically similar. A short, wide branch can take more than its share while a longer branch with extra bends gets less. That imbalance changes how much coolant reaches each GPU, even when the pump's total flow looks healthy.
- Fitting and bend losses. Every elbow, tee, coupler, and adapter adds a local pressure drop, and those small losses accumulate across a long run. Each one scales with the square of local velocity, so tight bends and narrow adapters hurt more at high flow. A neat, short hose route with few direction changes can free up head for the cold plates; a tangled route spends it on plumbing.
The combined effect is the system curve. It rises from the origin: at zero flow, no pressure is needed; as flow increases, the required pressure climbs. The pump operates where its own curve crosses that system curve. Change the number of plates, the hose length, or the branch layout, and the crossing point moves.
Why a nominal 1300 L/H pump still depends on the full loop design
The P90D pump used in the BC9 external water cooling system carries nominal maximum figures of 5 m head and 1300 L/H flow, with a nominal speed of 4500 RPM. Those are useful reference points, but they sit at opposite ends of the pump curve. The 1300 L/H number is a free-delivery figure: the flow a pump can reach when the outlet is essentially unrestricted. A real loop is never unrestricted. Every cold plate, radiator, hose, and bend adds resistance and pulls the operating point down from that ceiling. How far down depends on the full design. A short, simple loop with one cold plate and wide-bore tubing might run at a substantial fraction of the nominal flow. A larger multi-GPU loop with several plates, long hose runs, and many direction changes will sit lower on the curve. Coolant choice matters too: a glycol-rich mix is more viscous than water and raises pressure drop for the same flow. Trapped air has an outsized effect because it can stall sections of the loop and make the pump work against a moving pocket of gas rather than liquid. A common commissioning observation makes this concrete: add a second or third cold plate, or replace a short straight hose run with a longer route, and the measured flow drops even when the pump speed setting is unchanged. The tach signal may show the same 4500 RPM, because the impeller is still turning at that speed. RPM confirms rotation, not coolant delivery. A 4-pin tach interface is valuable for control and fault detection, but it is not a flow meter. The 1300 L/H figure is a nominal maximum; real system flow depends on the installed loop. To predict performance, compare the pump curve with an estimate of loop resistance, or measure flow directly after commissioning. Even a rough system curve helps: it shows whether adding a fourth GPU plate will merely trim flow or push the pump into a region where flow falls sharply. The operating temperature range of -10°C to 60°C and the 1.8 A current draw describe the pump's electrical and thermal envelope, while the hydraulic result still comes from the installed circuit.
Conclusion
Pump head and flow rate answer different questions. Head tells you how much pressure the pump can build to overcome the loop's resistance; flow tells you how much coolant actually moves. The two meet at a single operating point set by the pump curve and the system curve. In a multi-GPU loop, series plates, parallel branches, tubing, and bends all move that point, which is why a nominal 1300 L/H rating cannot describe every installation. Readers who want to see the exact published figures for the P90D pump can review the BC9 specifications and compare them with their own loop layout.
FAQ
Q: What is the difference between pump head and flow rate in a liquid cooling loop?
A: Head is pressure-raising capability, usually expressed in metres of liquid column; flow is volume moved per unit time, such as litres per hour. Head overcomes the pressure drop created by cold plates, tubing, and fittings. Flow is the result of pushing against that resistance. A pump has a curve that links the two, so maximum head and maximum flow occur at different operating points.
Q: Why does a multi-GPU loop need more pump head than a single cold plate loop?
A: More cold plates mean more pressure drop, especially in series where each plate's loss adds to the next. Longer tubing and more branches also add resistance. The pump must supply enough pressure to push the required flow through the whole path. A single-block loop has a shorter, lower-resistance path, so the same pump usually operates at a higher flow.
Q: Does a 1300 L/H pump rating guarantee the same flow in every cooling loop?
A: No. 1300 L/H is a nominal maximum flow, measured near free delivery. Real flow depends on total loop resistance, coolant viscosity, air content, and pump speed. A restrictive multi-GPU loop will run below that figure. The published number is a ceiling for comparing pumps, not a promise for a specific installation.
Sources / References
Pump Head and Pressure: Conversion, Calculations, and Charts
Pipe and Tube System Components - Minor (Dynamic) Loss Coefficients
ASHRAE Data Center Resources | Data Centers
No comments:
Post a Comment